Cross-Market Liquidity Stress Forecasting Using Explainable Temporal Fusion Networks Under Leakage-Safe Evaluation Protocols

Abstract

Liquidity stress in financial markets propagates rapidly across asset classes and geographies, posing systemic risks that challenge both private risk management and regulatory oversight. While machine learning models offer considerable promise for early warning systems, their deployment in cross-market settings is hampered by two interrelated problems: the opacity of many deep learning architectures, which undermines trust and regulatory acceptance, and the prevalence of data leakage in evaluation protocols, which produces overly optimistic performance estimates that fail to replicate in real-time deployment. This paper develops an explainable temporal fusion network (TFT) framework specifically designed for cross-market liquidity stress forecasting. The TFT architecture combines multi-horizon attention mechanisms with built-in interpretability components, including variable selection networks and temporal self-attention, enabling analysts to identify the primary drivers of predicted stress events. To address the second challenge, we propose a leakage-safe evaluation protocol that enforces strict temporal consistency through purging, embargoing, and combinatorial purged cross-validation. The framework is tested on a multi-asset dataset spanning equity, bond, and foreign exchange markets over a fifteen-year period. Results demonstrate that the TFT model achieves statistically significant improvements in stress prediction accuracy compared to baseline approaches, while the leakage-safe evaluation reduces false positive rates by over forty percent relative to naive walk-forward methods. We further examine structural trade-offs involving model complexity, interpretability, and computational sustainability, and discuss governance implications for deploying such systems within regulatory stress-testing frameworks. The findings underscore the necessity of embedding both explanatory mechanisms and rigorous evaluation discipline into the design of financial early warning infrastructures.